Attentive neural processes based on reliable inferences for industrial equipment anomaly detection

In modern manufacturing, assessing equipment conditions has become increasingly costly due to the complexity of industrial machinery. While data-driven methods have partially addressed this challenge, traditional approaches are often limited by their assumption of a simple Gaussian data distribution. This assumption fails in high-dimensional, complex industrial scenarios, where traditional models cannot capture the true data distribution, reducing their effectiveness. This paper introduces a reliable inference attentive neural process (RIANP) based on normalizing flows (NFs) and neural ordinary differential equations (NODEs), a method for detecting anomalies in industrial equipment. NFs replace the fixed prior assumption of the attentive neural process (ANP) with a learnable prior distribution, addressing sampling holes caused by unlearnable priors. Next, NODEs model the posterior distribution, enabling smoother alignment between the learnable prior and complex posterior distributions. A validation of two industrial anomaly detection cases shows that the RIANP achieves an average F1 score of 94.64 %, a 7.5 % improvement over the ANP, and an AUC of 96.5 %, representing a 12 % enhancement.